GPT-5.6's Sol model slashes coding costs by 50% while beating Anthropic's Claude

OpenAI just pulled back the curtain on GPT-5.6, and the story isn’t just about raw intelligence — it’s about making frontier models radically cheaper to run. The flagship GPT-5.6 Sol model, with max reasoning cranked up, now outperforms Anthropic’s Claude Fable 5 on the Artificial Analysis Coding Agent Index at less than half the cost. That’s a direct shot across the bow in the enterprise coding assistant war.

The real magic is happening under the hood. OpenAI detailed a three-layer efficiency overhaul spanning model training, inference, and what they call the “agentic harness” used by Codex and ChatGPT Work. The team trained GPT-5.6 to achieve more work per token, essentially teaching the model to take a more direct path through a task rather than meandering. But the most fascinating part might be the snake eating its own tail: GPT-5.6 Sol, running inside Codex, autonomously rewrote its own production GPU kernels. These are the low-level instructions that execute the model’s math. The result? A 20% reduction in end-to-end serving costs. To verify this self-written code, OpenAI leaned on an open-source tool called FpSan (Floating-Point Sanitizer).

Other infrastructure wins piled up. The team overhauled load balancing across their global fleet, using Sol to analyze production traffic and sniff out previously invisible inefficiencies in how requests get distributed across clusters and accelerators. They also invested in speculative decoding and caching strategies. None of these tweaks sounds revolutionary in isolation, but compounded across a user base of 1 billion people and over 2 million businesses, the savings are enormous. It’s how a model as capable as Terra can match GPT-5.5 on intelligence benchmarks at half the price, and Luna — their cheapest model — comes in at 80% less than Sol.

What’s clear is that OpenAI sees efficiency as survival. In a compute-constrained world where demand outpaces hardware supply, the company that serves the most tokens per dollar wins. By making their own model the primary tool for optimizing their stack, they’ve created a flywheel where smarter code literally makes the next generation of models cheaper to run. The question now is whether competitors can replicate this self-optimizing loop without access to the same level of infrastructure introspection.